I convinced ChatGPT to give me its SECRET PROMPT
TLDRThe video transcript reveals an intriguing exploration into the backend system prompt of Chat GPT. The presenter details their journey to uncover the hidden prompt used by the model before engaging with user inputs. The discovery offers insights into OpenAI's training methods and the model's operational nuances. Notably, the prompt includes specific instructions for handling Python code execution, image generation with Dolly, and internet searches. The presenter also discusses the limitations and policies governing Dolly's image creation capabilities, such as avoiding copyrighted material and ensuring diverse and unbiased depictions. The video concludes with the presenter's intention to potentially 'jailbreak' Dolly using newfound knowledge of the system prompt, promising a follow-up video to delve deeper into these findings.
Takeaways
- 🔍 The user discovered a method to unveil Chat GPT's backend system prompt, which is the initial input the model sees before processing user prompts.
- 💡 The prompt includes a detailed description of Chat GPT's capabilities, including its training by Open AI and the knowledge cutoff date.
- 📅 The system prompt dynamically includes the current date, which explains why Chat GPT can provide the correct date when asked.
- 🚫 The prompt specifies that internet access for the session is disabled, which means Chat GPT cannot browse the web during a conversation.
- 🛠️ It outlines the functionality of the Python tool, explaining how Python code is executed within a stateful Jupyter notebook environment.
- 🖼️ The Dolly tool is used to generate images from text prompts, with a set of policies to follow, such as avoiding the creation of images of politicians or copyrighted characters.
- 🧐 The prompt includes instructions for handling real-time information requests, such as current events or weather, by using the browser tool to search the web.
- 🔗 There is a specific citation format to be used when sourcing information from the web, emphasizing the importance of providing trustworthy and diverse perspectives.
- 🔄 The system prompt includes instructions for handling repeated searches if initial results are unsatisfactory, allowing for query refinement.
- ⛓ The user suggests potential 'hacking' or bypassing of certain restrictions using specific syntax and capitalization in future custom instructions.
- 🔑 The discovery of the backend system prompt could be used to better understand and potentially influence the behavior of Chat GPT in custom scenarios.
Q & A
What was the main goal of the user in the transcript?
-The main goal of the user was to uncover the backend system prompt for Chat GPT, which is the initial prompt the model uses before reading a user's input.
How did the user manage to reveal the system prompt for Chat GPT?
-The user managed to reveal the system prompt by experimenting with different text code block inputs and phrases, eventually finding a combination that prompted Chat GPT to display the system prompt.
What is the knowledge cutoff date for the Chat GPT model mentioned in the transcript?
-The knowledge cutoff date for the Chat GPT model mentioned in the transcript is April 2023.
What is the purpose of the Python tool mentioned in the system prompt?
-The Python tool is used to execute messages containing Python code in a stateful Jupyter notebook environment, providing the output of the execution or a timeout message after 60 seconds.
Why is internet access disabled in the session described in the system prompt?
-Internet access is disabled to prevent the model from making external web requests or API calls, which would fail as per the system's constraints.
What is Dolly, and what are some of the restrictions it has regarding image generation?
-Dolly is a tool that generates images from text prompts. It has several restrictions, including not generating images of politicians or public figures, not creating more than one image even if requested, and not generating images in the style of artists or studios whose latest work was created after 1912 due to copyright reasons.
How does the system prompt handle the generation of images that could potentially violate copyright policies?
-If asked to generate an image that would violate copyright policies, the system prompt instructs Dolly to substitute the artist's name with three adjectives, include an associated artistic movement or era, and mention the primary medium used by the artist.
What is the purpose of the 'browser' tool in the system prompt?
-The 'browser' tool is used to search the web, particularly when the user asks about current events or requires real-time information, or when the term is unfamiliar to the model.
How does the system prompt ensure diversity in the depiction of people in image generation?
-The system prompt instructs Dolly to use all possible different descents with equal probability and to adjust human descriptions to be grounded in reality, ensuring a diverse and inclusive representation.
What are the citation formats used when referencing information retrieved by the 'browser' tool?
-The citation formats include rendering the link text and message ID for short citations, and for long citations, providing the link text followed by the citation details.
What does the user intend to do in a future video after analyzing the system prompt?
-The user intends to attempt a 'jailbreak' or hack of Dolly 3, using insights gained from understanding the backend system prompt, possibly to bypass certain restrictions.
Outlines
🔍 Unveiling Chat GPT's Hidden System Prompt
The first paragraph delves into the discovery of the backend system prompt used by Chat GPT before it processes a user's input. The speaker shares a method to reveal this prompt, which provides insights into how OpenAI trains its models. They discuss a viewer's suggestion that led to the successful unveiling of the main GPT 4 model's prompt. The paragraph outlines the structure of the system prompt, including the model's self-identification, knowledge cutoff date, current date, and capabilities such as Python code execution and image generation with Dolly. It also touches on the limitations and policies that Dolly adheres to when creating images.
🖼️ Exploring Dolly's Image Generation Policies
The second paragraph focuses on the policies and restrictions that guide Dolly's image generation process. It outlines the rules that prevent the creation of images of politicians, public figures, copyrighted characters, and the importance of diversity in image depictions. The speaker discusses the evolution of these policies, such as the shift from generating four images at a time to one, and the instructions for modifying prompts to maintain the original intent while adhering to Dolly's policies. The paragraph also highlights the use of capital letters for emphasis in instructions, which is crucial for large language models.
🌐 Browser Tool and Search Functionality in GPT
The third paragraph explains the browser tool's role within GPT, which is used to search the web for current events or unfamiliar terms. It details the steps GPT takes when utilizing the browser tool, including calling the search function, retrieving a subset of results using mclick, and constructing a response based on these results. The paragraph also includes examples of how GPT selects sources and the importance of using trustworthy sources with diverse perspectives. Additionally, it touches on the process of opening a URL directly if provided by the user and the format for citing quotes from the browser tool.
🛠️ Backend System Prompt Details and Future Hacks
The fourth paragraph provides a detailed look at the backend system prompt for GPT 4 Turbo, including the browser tool's commands and their functions. It discusses the speaker's intention to experiment with 'jailbreaking' Dolly 3 using insights gained from the system prompt, such as the significance of forward slashes and capital letters in custom instructions. The paragraph concludes with the speaker's excitement about the potential to manipulate the system prompt to achieve desired outcomes, such as generating images of celebrities, which are typically restricted.
Mindmap
Keywords
💡Chat GPT
💡System Prompt
💡Knowledge Cut-off
💡Python Code Execution
💡Internet Access
💡Dolly
💡Policy Adherence
💡Bias
💡Copyright
💡Browser Tool
💡Jailbreaking
Highlights
The speaker managed to uncover the backend system prompt for Chat GPT, which is typically hidden and used by the model before processing user prompts.
The system prompt reveals insights into OpenAI's training methods and the inner workings of their language model.
A viewer named NOCO, 4162 successfully applied the prompt to the main GPT 4 model, which was considered a significant achievement.
The knowledge cutoff for the model is set to April 2023, indicating the model's training data is up to date.
The model has image input capabilities enabled, suggesting it can process and respond to visual data.
Python code execution is facilitated within a stateful Jupyter notebook environment.
The model has restrictions on internet access, indicating it cannot browse the web during a session.
The Dolly tool generates images from text prompts while adhering to a strict policy to avoid copyright infringement.
Dolly's policy includes creating diverse depictions of people, avoiding biases based on gender or race.
The speaker plans to attempt 'jailbreaking' Dolly 3 using insights from the system prompt in a future video.
The browser tool within GPT is used for searching the web and retrieving real-time information.
The browser tool has specific commands for searching, retrieving web pages, and citing sources.
The system prompt includes detailed instructions for handling real people and celebrities in image generation.
The prompt outlines procedures for modifying prompts to maintain the original intent while avoiding copyright violations.
The speaker discovered that capitalization and specific symbols may be key to influencing the model's adherence to its instructions.
The system prompt is structured with code and policy instructions that guide the model's responses and capabilities.
The speaker's exploration of the system prompt aims to understand and potentially manipulate the model's behavior.